NFTs are on a clear collision course with artificial intelligence, which will in theory unlock all sorts of new utilities, including AI-driven avatars that can hold conversations.
However, because few people understand how this type of AI works, it’s alarmingly easy to hoodwink buyers. Recently, a handful of so-called “smart” or “intelligent” NFT projects have begun minting digital nothingburgers, marketing them as groundbreaking “AI,” gaining endorsements from influential investors and breaking resale records. Soon, the market will be flooded with “AI NFTs,” the vast majority of which are absolutely, positively, NGMI (not going to make it).
Lauren Kunze is the CEO of Pandorabots, a leading chatbot platform that powers conversational AI applications for developers and global brands, and its subsidiary ICONIQ, which creates embodied AI like Kuki: a socialbot that has exchanged over a billion messages with 25 million human fans. Kunze excels at explaining AI and its real world applications to a general audience, including at conferences like TEDx, Mobile World Congress and SXSW, and for publications like TechCrunch, Quartz, and Venturebeat. She has also authored four novels, published by HarperCollins, and holds a degree from Harvard in Literature and Language, and Neuroscience.
Here are five red flags to watch for:
1) It’s a cheap deepfake
If you’ve yet to see a deepfake-powered moving portrait, the first time feels like some serious Harry Potter magic come-to-life. But this seeming wizardry is actually quite accessible to mere muggles through numerous free apps (e.g., TokkingHeads, MyHeritage, Wombo) that let you turn any image into deepfake content based on a source video. So if you are creating or already own NFTs and you want to make them move, you can do it yourself on the cheap. Anything beyond minimal movement like blinking or breathing may break the illusion due to inevitable incongruities between the image and its underlying source video. Thus if you see a moving NFT that isn’t moving much, it’s almost certainly using off-the-shelf deepfake software, and not some special secret AI sauce.
These cheap deepfakes are being used to advertise things like portraits of historical figures which, when plugged into GPT-3, could converse with you in realtime. The thing is, running a real-time conversation with an AI requires a 3D CG model. So if you stumble upon a deepfake advertised as capable of conversing in realtime, alarm bells should sound because deepfakes don’t really run autonomously. Rather, they rely on superimposing someone’s face on an actor in a video during post production (for example, Zuck on Trey Parker or Tom Cruise on a talented TikTok impersonator), or during a video call (say, Deep Elon Zoom-bombing using free software like Avatarify).
2) It uses GPT-3 or has other third-party dependencies
The Generative Pre-trained Transformer 3 (GPT-3) language model is amazing at generating words and images in the style of a specific person or genre given enough examples. Giant, generative language models – apart from being slow to reply and so expensive to train that GPT-3, trained pre-COVID, remains unaware the virus exists – are trained on the public internet, which is a bit akin to drinking from a public toilet. Human beings are also typically toilet-mouthed monsters when talking to software (and each other) on the internet, but even the human players of a GPT-3-powered role play game were disturbed when the algorithm inserted children into sexual scenarios. That is why OpenAI takes great care to highly restrict chatbot use cases – and why any NFT project purporting to use GPT-3, especially to power add-on “personalities,” should raise eyebrows.
After years in limited beta with a long waitlist, OpenAI just announced that its GPT-3 API is now generally available to developers. (Until and unless Microsoft – which funded and apparently exclusively licensed GPT-3 – closes “Open” AI.) So far, some upstarts with early access have been using GPT-3 to create and sell unique “personalities” for your NFT. But now that the API is truly open to all, the reality is that anyone can generate a GPT-3 powered “personality” by feeding it a few example texts. Whether OpenAI allows your particular use case to remain live is another story. Last summer, the Chronicle published a heartbreaking tale about a man who used a GPT-3 powered service to bring his dead fiancé back to life based on their texts. Months later, in a heart-rending twist, OpenAI shuttered the service without warning for violating its stringent terms; Joshua’s fiancé Jessica essentially died twice. GPT-3 NFTs are liable to meet the same fate.NFTs are ultimately about verified ownership. If your chatty “AI” NFT depends on third-party services that the seller neither owns, develops, nor controls, what exactly do you actually own?
3) AI is “on the roadmap” (and all over the marketing…
www.coindesk.com